kernel/events/hw_breakpoint.c

   1 // SPDX-License-Identifier: GPL-2.0+
   2 /*
   3  * Copyright (C) 2007 Alan Stern
   4  * Copyright (C) IBM Corporation, 2009
   5  * Copyright (C) 2009, Frederic Weisbecker <fweisbec@gmail.com>
   6  *
   7  * Thanks to Ingo Molnar for his many suggestions.
   8  *
   9  * Authors: Alan Stern <stern@rowland.harvard.edu>
  10  *          K.Prasad <prasad@linux.vnet.ibm.com>
  11  *          Frederic Weisbecker <fweisbec@gmail.com>
  12  */
  13
  14 /*
  15  * HW_breakpoint: a unified kernel/user-space hardware breakpoint facility,
  16  * using the CPU's debug registers.
  17  * This file contains the arch-independent routines.
  18  */
  19
  20 #include <linux/hw_breakpoint.h>
  21
  22 #include <linux/bug.h>
  23 #include <linux/cpu.h>
  24 #include <linux/export.h>
  25 #include <linux/init.h>
  26 #include <linux/irqflags.h>
  27 #include <linux/kdebug.h>
  28 #include <linux/kernel.h>
  29 #include <linux/mutex.h>
  30 #include <linux/notifier.h>
  31 #include <linux/percpu.h>
  32 #include <linux/rhashtable.h>
  33 #include <linux/sched.h>
  34 #include <linux/slab.h>
  35
  36 /*
  37  * Constraints data
  38  */
  39 struct bp_cpuinfo {
  40         /* Number of pinned cpu breakpoints in a cpu */
  41         unsigned int    cpu_pinned;
  42         /* tsk_pinned[n] is the number of tasks having n+1 breakpoints */
  43 #ifdef hw_breakpoint_slots
  44         unsigned int    tsk_pinned[hw_breakpoint_slots(0)];
  45 #else
  46         unsigned int    *tsk_pinned;
  47 #endif
  48         /* Number of non-pinned cpu/task breakpoints in a cpu */
  49         unsigned int    flexible; /* XXX: placeholder, see fetch_this_slot() */
  50 };
  51
  52 static DEFINE_PER_CPU(struct bp_cpuinfo, bp_cpuinfo[TYPE_MAX]);
  53
  54 static struct bp_cpuinfo *get_bp_info(int cpu, enum bp_type_idx type)
  55 {
  56         return per_cpu_ptr(bp_cpuinfo + type, cpu);
  57 }
  58
  59 /* Keep track of the breakpoints attached to tasks */
  60 static struct rhltable task_bps_ht;
  61 static const struct rhashtable_params task_bps_ht_params = {
  62         .head_offset = offsetof(struct hw_perf_event, bp_list),
  63         .key_offset = offsetof(struct hw_perf_event, target),
  64         .key_len = sizeof_field(struct hw_perf_event, target),
  65         .automatic_shrinking = true,
  66 };
  67
  68 static bool constraints_initialized __ro_after_init;
  69
  70 /* Gather the number of total pinned and un-pinned bp in a cpuset */
  71 struct bp_busy_slots {
  72         unsigned int pinned;
  73         unsigned int flexible;
  74 };
  75
  76 /* Serialize accesses to the above constraints */
  77 static DEFINE_MUTEX(nr_bp_mutex);
  78
  79 #ifdef hw_breakpoint_slots
  80 /*
  81  * Number of breakpoint slots is constant, and the same for all types.
  82  */
  83 static_assert(hw_breakpoint_slots(TYPE_INST) == hw_breakpoint_slots(TYPE_DATA));
  84 static inline int hw_breakpoint_slots_cached(int type)  { return hw_breakpoint_slots(type); }
  85 static inline int init_breakpoint_slots(void)           { return 0; }
  86 #else
  87 /*
  88  * Dynamic number of breakpoint slots.
  89  */
  90 static int __nr_bp_slots[TYPE_MAX] __ro_after_init;
  91
  92 static inline int hw_breakpoint_slots_cached(int type)
  93 {
  94         return __nr_bp_slots[type];
  95 }
  96
  97 static __init int init_breakpoint_slots(void)
  98 {
  99         int i, cpu, err_cpu;
 100
 101         for (i = 0; i < TYPE_MAX; i++)
 102                 __nr_bp_slots[i] = hw_breakpoint_slots(i);
 103
 104         for_each_possible_cpu(cpu) {
 105                 for (i = 0; i < TYPE_MAX; i++) {
 106                         struct bp_cpuinfo *info = get_bp_info(cpu, i);
 107
 108                         info->tsk_pinned = kcalloc(__nr_bp_slots[i], sizeof(int), GFP_KERNEL);
 109                         if (!info->tsk_pinned)
 110                                 goto err;
 111                 }
 112         }
 113
 114         return 0;
 115 err:
 116         for_each_possible_cpu(err_cpu) {
 117                 for (i = 0; i < TYPE_MAX; i++)
 118                         kfree(get_bp_info(err_cpu, i)->tsk_pinned);
 119                 if (err_cpu == cpu)
 120                         break;
 121         }
 122
 123         return -ENOMEM;
 124 }
 125 #endif
 126
 127 __weak int hw_breakpoint_weight(struct perf_event *bp)
 128 {
 129         return 1;
 130 }
 131
 132 static inline enum bp_type_idx find_slot_idx(u64 bp_type)
 133 {
 134         if (bp_type & HW_BREAKPOINT_RW)
 135                 return TYPE_DATA;
 136
 137         return TYPE_INST;
 138 }
 139
 140 /*
 141  * Report the maximum number of pinned breakpoints a task
 142  * have in this cpu
 143  */
 144 static unsigned int max_task_bp_pinned(int cpu, enum bp_type_idx type)
 145 {
 146         unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
 147         int i;
 148
 149         for (i = hw_breakpoint_slots_cached(type) - 1; i >= 0; i--) {
 150                 if (tsk_pinned[i] > 0)
 151                         return i + 1;
 152         }
 153
 154         return 0;
 155 }
 156
 157 /*
 158  * Count the number of breakpoints of the same type and same task.
 159  * The given event must be not on the list.
 160  */
 161 static int task_bp_pinned(int cpu, struct perf_event *bp, enum bp_type_idx type)
 162 {
 163         struct rhlist_head *head, *pos;
 164         struct perf_event *iter;
 165         int count = 0;
 166
 167         rcu_read_lock();
 168         head = rhltable_lookup(&task_bps_ht, &bp->hw.target, task_bps_ht_params);
 169         if (!head)
 170                 goto out;
 171
 172         rhl_for_each_entry_rcu(iter, pos, head, hw.bp_list) {
 173                 if (find_slot_idx(iter->attr.bp_type) == type &&
 174                     (iter->cpu < 0 || cpu == iter->cpu))
 175                         count += hw_breakpoint_weight(iter);
 176         }
 177
 178 out:
 179         rcu_read_unlock();
 180         return count;
 181 }
 182
 183 static const struct cpumask *cpumask_of_bp(struct perf_event *bp)
 184 {
 185         if (bp->cpu >= 0)
 186                 return cpumask_of(bp->cpu);
 187         return cpu_possible_mask;
 188 }
 189
 190 /*
 191  * Report the number of pinned/un-pinned breakpoints we have in
 192  * a given cpu (cpu > -1) or in all of them (cpu = -1).
 193  */
 194 static void
 195 fetch_bp_busy_slots(struct bp_busy_slots *slots, struct perf_event *bp,
 196                     enum bp_type_idx type)
 197 {
 198         const struct cpumask *cpumask = cpumask_of_bp(bp);
 199         int cpu;
 200
 201         for_each_cpu(cpu, cpumask) {
 202                 struct bp_cpuinfo *info = get_bp_info(cpu, type);
 203                 int nr;
 204
 205                 nr = info->cpu_pinned;
 206                 if (!bp->hw.target)
 207                         nr += max_task_bp_pinned(cpu, type);
 208                 else
 209                         nr += task_bp_pinned(cpu, bp, type);
 210
 211                 if (nr > slots->pinned)
 212                         slots->pinned = nr;
 213
 214                 nr = info->flexible;
 215                 if (nr > slots->flexible)
 216                         slots->flexible = nr;
 217         }
 218 }
 219
 220 /*
 221  * For now, continue to consider flexible as pinned, until we can
 222  * ensure no flexible event can ever be scheduled before a pinned event
 223  * in a same cpu.
 224  */
 225 static void
 226 fetch_this_slot(struct bp_busy_slots *slots, int weight)
 227 {
 228         slots->pinned += weight;
 229 }
 230
 231 /*
 232  * Add a pinned breakpoint for the given task in our constraint table
 233  */
 234 static void toggle_bp_task_slot(struct perf_event *bp, int cpu,
 235                                 enum bp_type_idx type, int weight)
 236 {
 237         unsigned int *tsk_pinned = get_bp_info(cpu, type)->tsk_pinned;
 238         int old_idx, new_idx;
 239
 240         old_idx = task_bp_pinned(cpu, bp, type) - 1;
 241         new_idx = old_idx + weight;
 242
 243         if (old_idx >= 0)
 244                 tsk_pinned[old_idx]--;
 245         if (new_idx >= 0)
 246                 tsk_pinned[new_idx]++;
 247 }
 248
 249 /*
 250  * Add/remove the given breakpoint in our constraint table
 251  */
 252 static int
 253 toggle_bp_slot(struct perf_event *bp, bool enable, enum bp_type_idx type,
 254                int weight)
 255 {
 256         const struct cpumask *cpumask = cpumask_of_bp(bp);
 257         int cpu;
 258
 259         if (!enable)
 260                 weight = -weight;
 261
 262         /* Pinned counter cpu profiling */
 263         if (!bp->hw.target) {
 264                 get_bp_info(bp->cpu, type)->cpu_pinned += weight;
 265                 return 0;
 266         }
 267
 268         /* Pinned counter task profiling */
 269         for_each_cpu(cpu, cpumask)
 270                 toggle_bp_task_slot(bp, cpu, type, weight);
 271
 272         if (enable)
 273                 return rhltable_insert(&task_bps_ht, &bp->hw.bp_list, task_bps_ht_params);
 274         else
 275                 return rhltable_remove(&task_bps_ht, &bp->hw.bp_list, task_bps_ht_params);
 276 }
 277
 278 __weak int arch_reserve_bp_slot(struct perf_event *bp)
 279 {
 280         return 0;
 281 }
 282
 283 __weak void arch_release_bp_slot(struct perf_event *bp)
 284 {
 285 }
 286
 287 /*
 288  * Function to perform processor-specific cleanup during unregistration
 289  */
 290 __weak void arch_unregister_hw_breakpoint(struct perf_event *bp)
 291 {
 292         /*
 293          * A weak stub function here for those archs that don't define
 294          * it inside arch/.../kernel/hw_breakpoint.c
 295          */
 296 }
 297
 298 /*
 299  * Constraints to check before allowing this new breakpoint counter:
 300  *
 301  *  == Non-pinned counter == (Considered as pinned for now)
 302  *
 303  *   - If attached to a single cpu, check:
 304  *
 305  *       (per_cpu(info->flexible, cpu) || (per_cpu(info->cpu_pinned, cpu)
 306  *           + max(per_cpu(info->tsk_pinned, cpu)))) < HBP_NUM
 307  *
 308  *       -> If there are already non-pinned counters in this cpu, it means
 309  *          there is already a free slot for them.
 310  *          Otherwise, we check that the maximum number of per task
 311  *          breakpoints (for this cpu) plus the number of per cpu breakpoint
 312  *          (for this cpu) doesn't cover every registers.
 313  *
 314  *   - If attached to every cpus, check:
 315  *
 316  *       (per_cpu(info->flexible, *) || (max(per_cpu(info->cpu_pinned, *))
 317  *           + max(per_cpu(info->tsk_pinned, *)))) < HBP_NUM
 318  *
 319  *       -> This is roughly the same, except we check the number of per cpu
 320  *          bp for every cpu and we keep the max one. Same for the per tasks
 321  *          breakpoints.
 322  *
 323  *
 324  * == Pinned counter ==
 325  *
 326  *   - If attached to a single cpu, check:
 327  *
 328  *       ((per_cpu(info->flexible, cpu) > 1) + per_cpu(info->cpu_pinned, cpu)
 329  *            + max(per_cpu(info->tsk_pinned, cpu))) < HBP_NUM
 330  *
 331  *       -> Same checks as before. But now the info->flexible, if any, must keep
 332  *          one register at least (or they will never be fed).
 333  *
 334  *   - If attached to every cpus, check:
 335  *
 336  *       ((per_cpu(info->flexible, *) > 1) + max(per_cpu(info->cpu_pinned, *))
 337  *            + max(per_cpu(info->tsk_pinned, *))) < HBP_NUM
 338  */
 339 static int __reserve_bp_slot(struct perf_event *bp, u64 bp_type)
 340 {
 341         struct bp_busy_slots slots = {0};
 342         enum bp_type_idx type;
 343         int weight;
 344         int ret;
 345
 346         /* We couldn't initialize breakpoint constraints on boot */
 347         if (!constraints_initialized)
 348                 return -ENOMEM;
 349
 350         /* Basic checks */
 351         if (bp_type == HW_BREAKPOINT_EMPTY ||
 352             bp_type == HW_BREAKPOINT_INVALID)
 353                 return -EINVAL;
 354
 355         type = find_slot_idx(bp_type);
 356         weight = hw_breakpoint_weight(bp);
 357
 358         fetch_bp_busy_slots(&slots, bp, type);
 359         /*
 360          * Simulate the addition of this breakpoint to the constraints
 361          * and see the result.
 362          */
 363         fetch_this_slot(&slots, weight);
 364
 365         /* Flexible counters need to keep at least one slot */
 366         if (slots.pinned + (!!slots.flexible) > hw_breakpoint_slots_cached(type))
 367                 return -ENOSPC;
 368
 369         ret = arch_reserve_bp_slot(bp);
 370         if (ret)
 371                 return ret;
 372
 373         return toggle_bp_slot(bp, true, type, weight);
 374 }
 375
 376 int reserve_bp_slot(struct perf_event *bp)
 377 {
 378         int ret;
 379
 380         mutex_lock(&nr_bp_mutex);
 381
 382         ret = __reserve_bp_slot(bp, bp->attr.bp_type);
 383
 384         mutex_unlock(&nr_bp_mutex);
 385
 386         return ret;
 387 }
 388
 389 static void __release_bp_slot(struct perf_event *bp, u64 bp_type)
 390 {
 391         enum bp_type_idx type;
 392         int weight;
 393
 394         arch_release_bp_slot(bp);
 395
 396         type = find_slot_idx(bp_type);
 397         weight = hw_breakpoint_weight(bp);
 398         WARN_ON(toggle_bp_slot(bp, false, type, weight));
 399 }
 400
 401 void release_bp_slot(struct perf_event *bp)
 402 {
 403         mutex_lock(&nr_bp_mutex);
 404
 405         arch_unregister_hw_breakpoint(bp);
 406         __release_bp_slot(bp, bp->attr.bp_type);
 407
 408         mutex_unlock(&nr_bp_mutex);
 409 }
 410
 411 static int __modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
 412 {
 413         int err;
 414
 415         __release_bp_slot(bp, old_type);
 416
 417         err = __reserve_bp_slot(bp, new_type);
 418         if (err) {
 419                 /*
 420                  * Reserve the old_type slot back in case
 421                  * there's no space for the new type.
 422                  *
 423                  * This must succeed, because we just released
 424                  * the old_type slot in the __release_bp_slot
 425                  * call above. If not, something is broken.
 426                  */
 427                 WARN_ON(__reserve_bp_slot(bp, old_type));
 428         }
 429
 430         return err;
 431 }
 432
 433 static int modify_bp_slot(struct perf_event *bp, u64 old_type, u64 new_type)
 434 {
 435         int ret;
 436
 437         mutex_lock(&nr_bp_mutex);
 438         ret = __modify_bp_slot(bp, old_type, new_type);
 439         mutex_unlock(&nr_bp_mutex);
 440         return ret;
 441 }
 442
 443 /*
 444  * Allow the kernel debugger to reserve breakpoint slots without
 445  * taking a lock using the dbg_* variant of for the reserve and
 446  * release breakpoint slots.
 447  */
 448 int dbg_reserve_bp_slot(struct perf_event *bp)
 449 {
 450         if (mutex_is_locked(&nr_bp_mutex))
 451                 return -1;
 452
 453         return __reserve_bp_slot(bp, bp->attr.bp_type);
 454 }
 455
 456 int dbg_release_bp_slot(struct perf_event *bp)
 457 {
 458         if (mutex_is_locked(&nr_bp_mutex))
 459                 return -1;
 460
 461         __release_bp_slot(bp, bp->attr.bp_type);
 462
 463         return 0;
 464 }
 465
 466 static int hw_breakpoint_parse(struct perf_event *bp,
 467                                const struct perf_event_attr *attr,
 468                                struct arch_hw_breakpoint *hw)
 469 {
 470         int err;
 471
 472         err = hw_breakpoint_arch_parse(bp, attr, hw);
 473         if (err)
 474                 return err;
 475
 476         if (arch_check_bp_in_kernelspace(hw)) {
 477                 if (attr->exclude_kernel)
 478                         return -EINVAL;
 479                 /*
 480                  * Don't let unprivileged users set a breakpoint in the trap
 481                  * path to avoid trap recursion attacks.
 482                  */
 483                 if (!capable(CAP_SYS_ADMIN))
 484                         return -EPERM;
 485         }
 486
 487         return 0;
 488 }
 489
 490 int register_perf_hw_breakpoint(struct perf_event *bp)
 491 {
 492         struct arch_hw_breakpoint hw = { };
 493         int err;
 494
 495         err = reserve_bp_slot(bp);
 496         if (err)
 497                 return err;
 498
 499         err = hw_breakpoint_parse(bp, &bp->attr, &hw);
 500         if (err) {
 501                 release_bp_slot(bp);
 502                 return err;
 503         }
 504
 505         bp->hw.info = hw;
 506
 507         return 0;
 508 }
 509
 510 /**
 511  * register_user_hw_breakpoint - register a hardware breakpoint for user space
 512  * @attr: breakpoint attributes
 513  * @triggered: callback to trigger when we hit the breakpoint
 514  * @context: context data could be used in the triggered callback
 515  * @tsk: pointer to 'task_struct' of the process to which the address belongs
 516  */
 517 struct perf_event *
 518 register_user_hw_breakpoint(struct perf_event_attr *attr,
 519                             perf_overflow_handler_t triggered,
 520                             void *context,
 521                             struct task_struct *tsk)
 522 {
 523         return perf_event_create_kernel_counter(attr, -1, tsk, triggered,
 524                                                 context);
 525 }
 526 EXPORT_SYMBOL_GPL(register_user_hw_breakpoint);
 527
 528 static void hw_breakpoint_copy_attr(struct perf_event_attr *to,
 529                                     struct perf_event_attr *from)
 530 {
 531         to->bp_addr = from->bp_addr;
 532         to->bp_type = from->bp_type;
 533         to->bp_len  = from->bp_len;
 534         to->disabled = from->disabled;
 535 }
 536
 537 int
 538 modify_user_hw_breakpoint_check(struct perf_event *bp, struct perf_event_attr *attr,
 539                                 bool check)
 540 {
 541         struct arch_hw_breakpoint hw = { };
 542         int err;
 543
 544         err = hw_breakpoint_parse(bp, attr, &hw);
 545         if (err)
 546                 return err;
 547
 548         if (check) {
 549                 struct perf_event_attr old_attr;
 550
 551                 old_attr = bp->attr;
 552                 hw_breakpoint_copy_attr(&old_attr, attr);
 553                 if (memcmp(&old_attr, attr, sizeof(*attr)))
 554                         return -EINVAL;
 555         }
 556
 557         if (bp->attr.bp_type != attr->bp_type) {
 558                 err = modify_bp_slot(bp, bp->attr.bp_type, attr->bp_type);
 559                 if (err)
 560                         return err;
 561         }
 562
 563         hw_breakpoint_copy_attr(&bp->attr, attr);
 564         bp->hw.info = hw;
 565
 566         return 0;
 567 }
 568
 569 /**
 570  * modify_user_hw_breakpoint - modify a user-space hardware breakpoint
 571  * @bp: the breakpoint structure to modify
 572  * @attr: new breakpoint attributes
 573  */
 574 int modify_user_hw_breakpoint(struct perf_event *bp, struct perf_event_attr *attr)
 575 {
 576         int err;
 577
 578         /*
 579          * modify_user_hw_breakpoint can be invoked with IRQs disabled and hence it
 580          * will not be possible to raise IPIs that invoke __perf_event_disable.
 581          * So call the function directly after making sure we are targeting the
 582          * current task.
 583          */
 584         if (irqs_disabled() && bp->ctx && bp->ctx->task == current)
 585                 perf_event_disable_local(bp);
 586         else
 587                 perf_event_disable(bp);
 588
 589         err = modify_user_hw_breakpoint_check(bp, attr, false);
 590
 591         if (!bp->attr.disabled)
 592                 perf_event_enable(bp);
 593
 594         return err;
 595 }
 596 EXPORT_SYMBOL_GPL(modify_user_hw_breakpoint);
 597
 598 /**
 599  * unregister_hw_breakpoint - unregister a user-space hardware breakpoint
 600  * @bp: the breakpoint structure to unregister
 601  */
 602 void unregister_hw_breakpoint(struct perf_event *bp)
 603 {
 604         if (!bp)
 605                 return;
 606         perf_event_release_kernel(bp);
 607 }
 608 EXPORT_SYMBOL_GPL(unregister_hw_breakpoint);
 609
 610 /**
 611  * register_wide_hw_breakpoint - register a wide breakpoint in the kernel
 612  * @attr: breakpoint attributes
 613  * @triggered: callback to trigger when we hit the breakpoint
 614  * @context: context data could be used in the triggered callback
 615  *
 616  * @return a set of per_cpu pointers to perf events
 617  */
 618 struct perf_event * __percpu *
 619 register_wide_hw_breakpoint(struct perf_event_attr *attr,
 620                             perf_overflow_handler_t triggered,
 621                             void *context)
 622 {
 623         struct perf_event * __percpu *cpu_events, *bp;
 624         long err = 0;
 625         int cpu;
 626
 627         cpu_events = alloc_percpu(typeof(*cpu_events));
 628         if (!cpu_events)
 629                 return (void __percpu __force *)ERR_PTR(-ENOMEM);
 630
 631         cpus_read_lock();
 632         for_each_online_cpu(cpu) {
 633                 bp = perf_event_create_kernel_counter(attr, cpu, NULL,
 634                                                       triggered, context);
 635                 if (IS_ERR(bp)) {
 636                         err = PTR_ERR(bp);
 637                         break;
 638                 }
 639
 640                 per_cpu(*cpu_events, cpu) = bp;
 641         }
 642         cpus_read_unlock();
 643
 644         if (likely(!err))
 645                 return cpu_events;
 646
 647         unregister_wide_hw_breakpoint(cpu_events);
 648         return (void __percpu __force *)ERR_PTR(err);
 649 }
 650 EXPORT_SYMBOL_GPL(register_wide_hw_breakpoint);
 651
 652 /**
 653  * unregister_wide_hw_breakpoint - unregister a wide breakpoint in the kernel
 654  * @cpu_events: the per cpu set of events to unregister
 655  */
 656 void unregister_wide_hw_breakpoint(struct perf_event * __percpu *cpu_events)
 657 {
 658         int cpu;
 659
 660         for_each_possible_cpu(cpu)
 661                 unregister_hw_breakpoint(per_cpu(*cpu_events, cpu));
 662
 663         free_percpu(cpu_events);
 664 }
 665 EXPORT_SYMBOL_GPL(unregister_wide_hw_breakpoint);
 666
 667 /**
 668  * hw_breakpoint_is_used - check if breakpoints are currently used
 669  *
 670  * Returns: true if breakpoints are used, false otherwise.
 671  */
 672 bool hw_breakpoint_is_used(void)
 673 {
 674         int cpu;
 675
 676         if (!constraints_initialized)
 677                 return false;
 678
 679         for_each_possible_cpu(cpu) {
 680                 for (int type = 0; type < TYPE_MAX; ++type) {
 681                         struct bp_cpuinfo *info = get_bp_info(cpu, type);
 682
 683                         if (info->cpu_pinned)
 684                                 return true;
 685
 686                         for (int slot = 0; slot < hw_breakpoint_slots_cached(type); ++slot) {
 687                                 if (info->tsk_pinned[slot])
 688                                         return true;
 689                         }
 690                 }
 691         }
 692
 693         return false;
 694 }
 695
 696 static struct notifier_block hw_breakpoint_exceptions_nb = {
 697         .notifier_call = hw_breakpoint_exceptions_notify,
 698         /* we need to be notified first */
 699         .priority = 0x7fffffff
 700 };
 701
 702 static void bp_perf_event_destroy(struct perf_event *event)
 703 {
 704         release_bp_slot(event);
 705 }
 706
 707 static int hw_breakpoint_event_init(struct perf_event *bp)
 708 {
 709         int err;
 710
 711         if (bp->attr.type != PERF_TYPE_BREAKPOINT)
 712                 return -ENOENT;
 713
 714         /*
 715          * no branch sampling for breakpoint events
 716          */
 717         if (has_branch_stack(bp))
 718                 return -EOPNOTSUPP;
 719
 720         err = register_perf_hw_breakpoint(bp);
 721         if (err)
 722                 return err;
 723
 724         bp->destroy = bp_perf_event_destroy;
 725
 726         return 0;
 727 }
 728
 729 static int hw_breakpoint_add(struct perf_event *bp, int flags)
 730 {
 731         if (!(flags & PERF_EF_START))
 732                 bp->hw.state = PERF_HES_STOPPED;
 733
 734         if (is_sampling_event(bp)) {
 735                 bp->hw.last_period = bp->hw.sample_period;
 736                 perf_swevent_set_period(bp);
 737         }
 738
 739         return arch_install_hw_breakpoint(bp);
 740 }
 741
 742 static void hw_breakpoint_del(struct perf_event *bp, int flags)
 743 {
 744         arch_uninstall_hw_breakpoint(bp);
 745 }
 746
 747 static void hw_breakpoint_start(struct perf_event *bp, int flags)
 748 {
 749         bp->hw.state = 0;
 750 }
 751
 752 static void hw_breakpoint_stop(struct perf_event *bp, int flags)
 753 {
 754         bp->hw.state = PERF_HES_STOPPED;
 755 }
 756
 757 static struct pmu perf_breakpoint = {
 758         .task_ctx_nr    = perf_sw_context, /* could eventually get its own */
 759
 760         .event_init     = hw_breakpoint_event_init,
 761         .add            = hw_breakpoint_add,
 762         .del            = hw_breakpoint_del,
 763         .start          = hw_breakpoint_start,
 764         .stop           = hw_breakpoint_stop,
 765         .read           = hw_breakpoint_pmu_read,
 766 };
 767
 768 int __init init_hw_breakpoint(void)
 769 {
 770         int ret;
 771
 772         ret = rhltable_init(&task_bps_ht, &task_bps_ht_params);
 773         if (ret)
 774                 return ret;
 775
 776         ret = init_breakpoint_slots();
 777         if (ret)
 778                 return ret;
 779
 780         constraints_initialized = true;
 781
 782         perf_pmu_register(&perf_breakpoint, "breakpoint", PERF_TYPE_BREAKPOINT);
 783
 784         return register_die_notifier(&hw_breakpoint_exceptions_nb);
 785 }